Decisions that put the record's meaning before the deadline

Ending row-level extracts and losing a €1.4 million study

In 2022 Aurevalle sent 21 pseudonymised extracts of patient data to research partners: universities, two cooperative cancer groups and several pharmaceutical sponsors. Each left under a data-use agreement and a promise of deletion at the end of the project. In October that year one extract, covering 1,180 patients, was emailed to the wrong researcher at a university partner. We notified the CNIL within 72 hours and the file was deleted, but the episode made me ask how many copies of our patients' records now existed outside the group. Nobody could tell me.

In 2023 a pharmaceutical sponsor proposed a three-year, €1.4 million real-world study of 9,200 lung cancer patients and asked for an extract to be sent to its analytics contractor. I recommended that we stop sending row-level extracts to any partner and build an environment in which external analysts would work on our data under our control, with only checked aggregate results leaving it. Our director of research and innovation, a respected oncologist, opposed me. Partnership income, then €1.9 million a year, paid for part of his research staff, and he expected academic co-authors to resent the constraint. The sponsor's evidence lead said its validated pipelines could run only on its own servers. The sponsor withdrew.

The executive committee backed the change in September 2023, and the foundation board approved €2.1 million of the warehouse programme for the environment, which opened in April 2024. Two trained stewards check every output before release, suppressing any cell below ten patients, and every session is logged. Income fell to €1.3 million in 2024, and two further sponsors left because the environment did not yet support the software versions their procedures required.

In 2025, 38 external projects ran inside the environment, 24 of them academic, and partnership income reached €2.6 million. No row-level patient record has left the group since April 2024. One output check stopped a table that would have identified three patients with a rare sarcoma subtype in a small town.

I would change two things. I would have invited sponsors' analysts to test the environment before it opened, which would probably have kept those two sponsors. And I would have explained the change to our own clinician-researchers before announcing it to partners; several first heard about it from a sponsor.

Keeping a chest X-ray algorithm out of intensive care

By late 2023 our radiologists were reporting some chest radiographs two or three days after they were taken at Valenciennes and Dunkirk, where three radiologist posts had been vacant for more than a year. The executive committee agreed to buy a CE-marked algorithm that flags suspected pneumothorax, pleural effusion and lung nodules, and our chief of radiology wanted it live at all six hospitals by March 2024. We take about 290,000 chest radiographs a year. A regional innovation grant would pay €180,000 if the tool went live by the end of June.

The vendor's study covered more than 40,000 images and reported high sensitivity. I asked for four months of silent running first: the algorithm would read our films with its output hidden from clinicians, and we would compare it with the radiologists' reports, with a panel of three radiologists adjudicating every disagreement. The chief of radiology argued, with reason, that patients were waiting now and that the device was already certified. The vendor called local testing unnecessary. Nothing in its study resembled our portable intensive-care films.

Silent running covered 26,400 films at three hospitals and cost €310,000, mostly radiologists' panel time. On upright adult films from emergency and outpatient departments, sensitivity for pneumothorax was 0.94, close to the vendor's figure. On supine portable films from intensive care it was 0.68, and tubes and drains caused so many false alarms that only about one flag in five was correct. Children were outside the product's intended use and had been arriving in its queue anyway.

We went live in July 2024 for upright adult films in emergency and outpatient settings at all six hospitals, excluding intensive care and children, and we lost the grant payment. In 2025 the tool read 182,000 films, and the median time to report a flagged urgent film at Valenciennes and Dunkirk fell from 31 hours to four. Monthly monitoring caught a drop in specificity after a scanner replacement at Arras in November 2025; we suspended the flag there for five weeks until radiographers corrected the new unit's processing settings.

What I would do differently concerns people rather than statistics. Intensive-care physicians learnt of their exclusion from a memo and read it as a verdict on their unit. I should have shown them their own results first.

Postponing a database lock over blood pressure readings that looked too regular

In 2014 Orsinval was running a phase III trial for a European sponsor's antihypertensive medicine in patients with chronic kidney disease: 2,400 patients at 160 sites in eleven countries, a twenty-month treatment period and database lock planned six weeks after the last patient's final visit. The sponsor was our second-largest client, and our contract paid a €650,000 milestone if the database locked on the agreed date.

In month fourteen, a statistician in my team running exploratory checks noticed that one high-recruiting site, with 74 patients, reported blood pressure readings whose visit-to-visit variability was far lower than at any other site, with a strong preference for values ending in zero and heart rates that repeated from one patient to the next. None of it broke an edit check. Each value was plausible on its own.

I recommended telling the sponsor at once, proposing a for-cause audit and planning for a late lock. Our account director disagreed. The site held about three per cent of patients, a sensitivity analysis excluding it could be run after lock, and raising an unproven suspicion of fabrication with our second-largest client seemed to him reckless. I argued that the decision belonged to the sponsor, that reopening a locked database is far worse than delaying it, and that a contract research organisation which sits on such a pattern has made the sponsor's choice for it. Our chief operating officer supported me.

The audit, run by the sponsor's quality team with our quality assurance group, found that visit data for 41 of the 74 patients had been fabricated or copied from other records. The sponsor excluded the site, reported the findings as good clinical practice requires and locked the database ten weeks late. We forfeited the €650,000 milestone. The primary result held: without the site, systolic blood pressure fell by 6.1 mmHg more than with placebo, against 6.8 mmHg with the site included.

From 2015 Orsinval ran central statistical monitoring on every phase II and III study from the first patient onward, and two years later the same sponsor awarded us its next phase III programme.

What I would do differently is simple: run those checks from the first month. At month three the site had enrolled twelve patients, and we would have found the pattern before most of the fabricated visits were recorded.

Withdrawing an external control arm from a rare-disease dossier

In 2019 Tornvig Pharma prepared reimbursement dossiers in four European countries for a medicine treating a rare inherited lung disease. It had been approved on a single-arm trial of 38 patients, which showed lung function declining more slowly than expected. To quantify the benefit for health technology assessment bodies, our market access team planned to compare the trial with patients from a European natural-history registry run by an academic consortium.

My team's feasibility review found three problems. The registry recorded lung function at a median interval of fourteen months, against three months in the trial. Baseline measurements were missing for 46 per cent of its 212 patients. And spirometry methods varied between centres in ways the registry did not capture. After matching on age and severity, 29 comparators remained, measured too rarely to estimate a rate of decline with any confidence.

I recommended presenting the registry descriptively, as context, and withdrawing it as the primary comparison. I also proposed that Tornvig fund standardised data collection within the registry, €2.4 million over three years, so that a future reassessment could rest on comparable measurements. Our head of market access for Europe disagreed strongly. He argued that every assessment body would ask for a comparison, that an imperfect one was better than none, and that price negotiations in two countries depended on it. He was not wrong about the risk. The senior vice president for global medical affairs supported my recommendation, and the executive committee approved it after I presented it to the board's science committee.

The cost was real. In two of the four countries, the first assessments did not accept that the added benefit could be quantified, and in one the negotiated price ended about 15 per cent below the company's plan. In the other two, assessors noted the limitations but accepted the trial on its own terms. By the time I left in early 2021, eleven centres were collecting lung function to a common protocol and 140 patients had been enrolled.

What changed was the order of work: every Tornvig evidence plan from 2020 began with a fitness assessment of any external data before a trial's design was fixed. That is also what I would do differently. The mismatch was visible before the single-arm trial began, and nobody, including me, had asked the question at that stage.